An array of semiconducting thin-film sensors is used for high-accuracy determination of gas concentrations in a binary mixture of H2S and NO2. For this purpose we developed a newly conceived pre-processing method for data feeding of a standard neural network. Pre-processing exploits dimension compaction of the vectorial signal of the sensor array. This operation is being accomplished by an auxiliary auto-associative network, separately trained from the main one. Our results present unprecedented accuracy even when applied to a validation-data set and appear to be of widespread application.